Files
b2txt25/language_model/srilm-1.7.3/flm/test/tests/ngram-factored/test.flm
2025-07-02 12:18:09 -07:00

77 lines
2.1 KiB
Plaintext

##
##
##
## Factord Language Model File:
##
## Current set of Node Options
##
## gtmin [num]
## gtmax [num]
## gt [fileName string]
## cdiscount [double]
## ndiscount []
## wbdiscount []
## kndiscount []
## ukndiscount []
## kn-counts-modified []
## kn-counts-modify-at-end []
## kn [fileName string]
## kn-count-parent [parent spec]
## interpolate []
## write [fileName string]
##
## ## The next set of options are active only when there
## ## are multiple backoff paths (backoff-graph children) possible.
##
## strategy [option]
## where [option] is one of:
## counts_no_norm
## counts_sum_counts_norm <default>
## counts_sum_num_words_norm
## counts_prod_card_norm
## counts_sum_card_norm
## counts_sum_log_card_norm
## bog_node_prob
## combine [option]
## where [option] is one of:
## max <default>
## min
## sum
## avg||mean
## prod
## gmean
## wmean { <node_spec weight> <node_spec weight> ... }
##
##
##
## Factors that are currently available in the files in this directory:
##
## W - word (about 14k)
## M - morphological class (about 1279)
## S - stem (about 5281)
## R - root (about 3346)
## P - pattern (about 1516)
##
1
## bigram w. general backoff that gets better than trigram.
## logprob= -84967 ppl= 168.703 ppl1= 439.536
W : 3 W(-1) M(-1) S(-1) dev.count.gz dev.lm.gz 5
W1,M1,S1 W1 kndiscount gtmin 2 interpolate
M1,S1 S1 kndiscount gtmin 2 interpolate
M1 M1 kndiscount gtmin 2 interpolate
S1 S1 kndiscount gtmin 1 interpolate
0 0 kndiscount gtmin 1
## bigram w. general backoff that gets better than trigram.
## logprob= -84967 ppl= 168.703 ppl1= 439.536
W : 3 W(-1) M(-1) S(-1) dev.count.gz dev.lm.gz 5
W1,M1,S1 W1 kndiscount gtmin 2 interpolate
M1,S1 S1,M1 kndiscount gtmin 100000000 combine mean
M1 M1 kndiscount gtmin 3 kn-count-parent W1,M1,S1
S1 S1 kndiscount gtmin 1 kn-count-parent W1,M1,S1
0 0 kndiscount gtmin 1 kn-count-parent W1,M1,S1